Missing architectures jsonnet files

[jeremiahvpratt]

ma_policy/graph_construct.py specifies that file mas/ppo/base-architectures.jsonnet contains example architectures, to the best of my ability I can't find that file in the repository.

[todor-markov]

My bad, that specific comment was leftover from the internal version of the file - removed now. Currently, if you want to see the architecture for a given policy, you should load the .npz file for that policy and extract it from there:

policy_dict = dict(np.load(policy_path))
policy_fn_and_args_raw = pickle.loads(policy_dict['policy_fn_and_args'])
policy_args = policy_fn_and_args_raw['args']  # this contains the arguments fed in to the policy class
network_spec = policy_args['network_spec']  # this contains the network architecture
print(network_spec)

I believe all of the policies we provide have the same architecture we describe in the paper.

[phiresky]

Just to make it easier for others (since the code depends on an old gym version and the paper is somewhat underspecified), here's the json for hide_and_seek_full :

[
  {
    "activation": "relu",
    "filters": 9,
    "kernel_size": 3,
    "layer_type": "circ_conv1d",
    "nodes_in": ["lidar"],
    "nodes_out": ["lidar"]
  },
  {
    "layer_type": "flatten_outer",
    "nodes_in": ["lidar"],
    "nodes_out": ["lidar"]
  },
  {
    "layer_type": "concat",
    "nodes_in": ["main", "lidar"],
    "nodes_out": ["main"]
  },
  {
    "layer_type": "concat",
    "nodes_in": ["main", "agent_qpos_qvel"],
    "nodes_out": ["agent_qpos_qvel"]
  },
  {
    "layer_type": "concat",
    "nodes_in": ["main", "box_obs"],
    "nodes_out": ["box_obs"]
  },
  {
    "layer_type": "concat",
    "nodes_in": ["main", "ramp_obs"],
    "nodes_out": ["ramp_obs"]
  },
  {
    "activation": "relu",
    "layer_type": "dense",
    "nodes_in": ["agent_qpos_qvel", "box_obs", "ramp_obs", "main"],
    "nodes_out": ["agent_qpos_qvel", "box_obs", "ramp_obs", "main"],
    "units": 128
  },
  {
    "layer_type": "entity_concat",
    "mask_out": "objects_mask",
    "masks_in": ["mask_aa_obs", "mask_ab_obs", "mask_ar_obs", None],
    "nodes_in": ["agent_qpos_qvel", "box_obs", "ramp_obs", "main"],
    "nodes_out": ["objects"]
  },
  {
    "heads": 4,
    "layer_norm": True,
    "layer_type": "residual_sa_block",
    "mask": "objects_mask",
    "n_embd": 128,
    "n_mlp": 1,
    "nodes_in": ["objects"],
    "nodes_out": ["objects"],
    "post_sa_layer_norm": True
  },
  {
    "layer_type": "entity_pooling",
    "mask": "objects_mask",
    "nodes_in": ["objects"],
    "nodes_out": ["objects_pooled"]
  },
  {
    "layer_type": "concat",
    "nodes_in": ["main", "objects_pooled"],
    "nodes_out": ["main"]
  },
  { "layer_type": "layernorm" },
  { "activation": "relu", "layer_type": "dense", "units": 256 },
  { "layer_type": "layernorm" },
  { "layer_type": "lstm", "units": 256 },
  { "layer_type": "layernorm" }
]

script:

import numpy as np
import pickle
from pprint import pprint
policy_path = "examples/hide_and_seek_full.npz"
policy_dict = dict(np.load(policy_path))
policy_fn_and_args_raw = pickle.loads(policy_dict['policy_fn_and_args'])
policy_args = policy_fn_and_args_raw['args']  # this contains the arguments fed in to the policy class
network_spec = policy_args['network_spec']  # this contains the network architecture
pprint(network_spec)